claude-atlas

---

name: claude-atlas description: Meta-skill for evolving the claude-atlas project itself. Use this skill whenever working inside the claude-atlas repository — adding new artifact kinds (agent/skill/command/memory + future ones), adding new edge detectors (duplicate, override, collision, reference + future ones), tuning heuristic thresholds (Jaccard, trigger overlap), extending the HTML report, or touching the Anthropic SDK LLM-judge pass. Triggers on: "claude-atlas", "atlas graph", "artifact detector", "edge kind", "jaccard threshold", "semantic duplicate", "trigger collision", "report template", "cytoscape", or any work inside this repo's src/claude_atlas/ tree.

claude-atlas meta-skill

You are working on claude-atlas, a CLI that scans a user's Claude Code setup and renders a relationship graph. This skill tells you the invariants that must survive every change.

When to use

• Adding new ArtifactKind values (e.g., hooks, plugins).
• Adding new EdgeKind values or new detectors in analysis/graph.py.
• Tuning thresholds (JACCARD_SUSPICIOUS, JACCARD_DUPLICATE, TRIGGER_COLLISION_MIN_SHARED).
• Evolving the HTML report (report/renderer.py + report/templates/report.mustache).
• Touching the LLM-as-judge pass in analysis/llm_judge.py.

Core invariants

1. Heuristics are deterministic and LLM-free by default. The --semantic flag is opt-in. Never make network I/O a default. The CLI must be usable offline.
2. Detectors are pure functions. detect_* in graph.py take list[Artifact] and return list[Edge]. No side effects, no I/O, no global state. Makes them trivially testable.
3. Artifact IDs are stable and unique. The current recipe is "{kind}:{scope}:{name}:{path}". If you change it, update every detector that builds edges by ID.
4. The HTML report is self-contained. Cytoscape is loaded via CDN, but there's no other runtime network dependency. Don't add fetches to the template.
5. PT-BR + EN docs are mandatory. Every user-facing doc change goes in both README.md and README.pt-BR.md. Keep them structurally identical.
6. Copyright / ownership: every new file Gabriel's tooling creates stays MIT. Never add a non-OSI-approved header.

Adding a new edge detector (pattern)

1. Add the enum value to EdgeKind in models.py.
2. If it should appear in the "Issues" tab, add it to _ISSUE_KINDS in report/renderer.py and to ScanResult.issues in models.py.
3. Write a detect_<thing>(artifacts: list[Artifact]) -> list[Edge] in analysis/graph.py. Keep it pure.
4. Call it from build_all_edges.
5. Pick a color in _edge_color (report/renderer.py) and add a legend entry in the Mustache template.
6. Write at least one focused test in tests/test_core.py that constructs artifacts by hand and asserts exactly the edges you expect.

Adding a new artifact kind

1. Add to ArtifactKind enum.
2. Extend scan_claude_dir in scanner/parsers.py to find it.
3. Add a color in _node_color and a legend dot in the template.
4. Decide whether existing detectors should include it. Default: most detectors skip MEMORY — follow that precedent unless there's a reason not to.

Threshold changes

Thresholds live at the top of analysis/graph.py as module constants. Never hardcode them inside detector bodies — keep them discoverable. If you add a new threshold, mention it in both READMEs' "What it detects" section.

LLM judge etiquette

• llm_judge.py imports anthropic lazily inside refine_with_llm. Never add a top-level import.
• The default model is pinned to DEFAULT_MODEL (currently claude-sonnet-4-6). The CLI already exposes --model for overrides — don't bypass it.
• The judge only ever re-examines DUPLICATE_SEMANTIC edges. Don't widen its scope without updating docs.
• Output must stay structured JSON. If you extend the verdict schema, update both the prompt and the parser together.

Testing discipline

• pytest must pass on every commit. The suite is cheap; run it.
• ruff check . and mypy are soft-gated (not strict), but new code should not introduce new warnings.
• Prefer constructing Artifact objects directly in tests (see _make_artifact in test_core.py) rather than writing files to disk — fewer moving parts.